def train(opt):
if torch.cuda.is_available():
torch.cuda.manual_seed(123)
else:
torch.manual_seed(123)
learning_rate_schedule = {"0": 1e-5, "5": 1e-4, "80": 1e-5, "110": 1e-6}
twoShapesAnchors = [(1.0, 1.0), (3.0, 3.0), (6, 6), (9.0, 9.0),
(10.0, 5.0), (5.0, 10.0)]
twoShapesAnchors = [(1.3221, 1.73145), (3.19275, 4.00944),
(5.05587, 8.09892), (9.47112, 4.84053),
(11.2364, 10.0071)]
training_params = {
"batch_size": opt.batch_size,
"shuffle": True,
"drop_last": True,
"collate_fn": custom_collate_fn
}
test_params = {
"batch_size": opt.batch_size,
"shuffle": False,
"drop_last": False,
"collate_fn": custom_collate_fn
}
training_set = TwoShapesDataset(root_path=opt.data_path,
mode=opt.train_set,
trainingSet="annotations",
image_size=opt.image_size)
training_generator = DataLoader(training_set, **training_params)
test_set = TwoShapesDataset(root_path=opt.data_path,
mode=opt.test_set,
trainingSet="annotations",
image_size=opt.image_size,
is_training=False)
test_generator = DataLoader(test_set, **test_params)
if torch.cuda.is_available():
if opt.pre_trained_model_type == "model":
model = torch.load(opt.pre_trained_model_path)
elif opt.pre_trained_model_type == "params":
model = Yolo(training_set.num_classes, twoShapesAnchors)
model.load_state_dict(torch.load(opt.pre_trained_model_path))
else:
print("Just loading the model definintion")
model = Yolo(training_set.num_classes, twoShapesAnchors)
else:
if opt.pre_trained_model_type == "model":
model = torch.load(opt.pre_trained_model_path,
map_location=lambda storage, loc: storage)
elif opt.pre_trained_model_type == "params":
model = Yolo(training_set.num_classes, twoShapesAnchors)
model.load_state_dict(
torch.load(opt.pre_trained_model_path,
map_location=lambda storage, loc: storage))
# The following line will re-initialize weight for the last layer, which is useful
# when you want to retrain the model based on my trained weights. if you uncomment it,
# you will see the loss is already very small at the beginning.
# nn.init.normal_(list(model.modules())[:].weight, 0, 0.01)
# model.apply(weights_init)
log_path = os.path.join(opt.log_path, "{}".format("test1"))
if os.path.isdir(log_path):
shutil.rmtree(log_path)
os.makedirs(log_path)
writer = SummaryWriter(log_path)
if torch.cuda.is_available():
writer.add_graph(
model.cpu(),
torch.rand(opt.batch_size, 3, opt.image_size, opt.image_size))
model.cuda()
else:
writer.add_graph(
model, torch.rand(opt.batch_size, 3, opt.image_size,
opt.image_size))
criterion = YoloLoss(training_set.num_classes, model.anchors,
opt.reduction)
optimizer = torch.optim.SGD(model.parameters(),
lr=1e-5,
momentum=opt.momentum,
weight_decay=opt.decay)
best_loss = 1e10
best_epoch = 0
model.train()
num_iter_per_epoch = len(training_generator)
for epoch in range(opt.num_epoches):
if str(epoch) in learning_rate_schedule.keys():
for param_group in optimizer.param_groups:
param_group['lr'] = learning_rate_schedule[str(epoch)]
for iter, batch in enumerate(training_generator):
image, label = batch
if torch.cuda.is_available():
image = Variable(image.cuda(), requires_grad=True)
else:
image = Variable(image, requires_grad=True)
optimizer.zero_grad()
logits = model(image)
loss, loss_coord, loss_conf, loss_cls = criterion(logits, label)
loss.backward()
optimizer.step()
print(
"Epoch: {}/{}, Iteration: {}/{}, Lr: {}, Loss:{:.2f} (Coord:{:.2f} Conf:{:.2f} Cls:{:.2f})"
.format(epoch + 1, opt.num_epoches, iter + 1,
num_iter_per_epoch, optimizer.param_groups[0]['lr'],
loss, loss_coord, loss_conf, loss_cls))
writer.add_scalar('Train/Total_loss', loss,
epoch * num_iter_per_epoch + iter)
writer.add_scalar('Train/Coordination_loss', loss_coord,
epoch * num_iter_per_epoch + iter)
writer.add_scalar('Train/Confidence_loss', loss_conf,
epoch * num_iter_per_epoch + iter)
writer.add_scalar('Train/Class_loss', loss_cls,
epoch * num_iter_per_epoch + iter)
if epoch % opt.test_interval == 0:
model.eval()
loss_ls = []
loss_coord_ls = []
loss_conf_ls = []
loss_cls_ls = []
for te_iter, te_batch in enumerate(test_generator):
te_image, te_label = te_batch
num_sample = len(te_label)
if torch.cuda.is_available():
te_image = te_image.cuda()
with torch.no_grad():
te_logits = model(te_image)
batch_loss, batch_loss_coord, batch_loss_conf, batch_loss_cls = criterion(
te_logits, te_label)
loss_ls.append(batch_loss * num_sample)
loss_coord_ls.append(batch_loss_coord * num_sample)
loss_conf_ls.append(batch_loss_conf * num_sample)
loss_cls_ls.append(batch_loss_cls * num_sample)
te_loss = sum(loss_ls) / test_set.__len__()
te_coord_loss = sum(loss_coord_ls) / test_set.__len__()
te_conf_loss = sum(loss_conf_ls) / test_set.__len__()
te_cls_loss = sum(loss_cls_ls) / test_set.__len__()
print(
"Epoch: {}/{}, Lr: {}, Loss:{:.2f} (Coord:{:.2f} Conf:{:.2f} Cls:{:.2f})"
.format(epoch + 1, opt.num_epoches,
optimizer.param_groups[0]['lr'], te_loss,
te_coord_loss, te_conf_loss, te_cls_loss))
writer.add_scalar('Test/Total_loss', te_loss, epoch)
writer.add_scalar('Test/Coordination_loss', te_coord_loss, epoch)
writer.add_scalar('Test/Confidence_loss', te_conf_loss, epoch)
writer.add_scalar('Test/Class_loss', te_cls_loss, epoch)
model.train()
if te_loss + opt.es_min_delta < best_loss:
best_loss = te_loss
best_epoch = epoch
# torch.save(model, opt.saved_path + os.sep + "trained_yolo_coco")
torch.save(
model.state_dict(),
opt.saved_path + os.sep + "only_params_trained_yolo_coco")
torch.save(model, opt.saved_path + os.sep + opt.save_file)
# Early stopping
if epoch - best_epoch > opt.es_patience > 0:
print(
"Stop training at epoch {}. The lowest loss achieved is {}"
.format(epoch, te_loss))
break
writer.export_scalars_to_json(log_path + os.sep + "all_logs.json")
writer.close()
def train(opt):
if not os.path.isdir(opt.saved_path):
os.makedirs(opt.saved_path)
os.environ['CUDA_VISIBLE_DEVICES'] = opt.gpu_devices
if torch.cuda.is_available():
torch.cuda.manual_seed(123)
else:
torch.manual_seed(123)
learning_rate_schedule = {"0": 1e-5, "5": 1e-4, "80": 1e-5, "110": 1e-6}
training_params = {
"batch_size": opt.batch_size,
"shuffle": True,
"drop_last": True,
"collate_fn": custom_collate_fn
}
test_params = {
"batch_size": opt.batch_size,
"shuffle": False,
"drop_last": False,
"collate_fn": custom_collate_fn
}
training_set = BOTDataset_scene(opt.data_path,
opt.train_set,
opt.image_size,
scene=2)
# mmm = training_set[0]
# print(1)
training_generator = DataLoader(training_set, **training_params)
test_set = BOTDataset_scene(opt.data_path,
opt.test_set,
opt.image_size,
scene=2,
is_training=False)
test_generator = DataLoader(test_set, **test_params)
if torch.cuda.is_available():
if opt.pre_trained_model_type == "model":
model = torch.load(opt.pre_trained_model_path)
else:
# model = Yolo(training_set.num_classes)
# model.load_state_dict(torch.load(opt.pre_trained_model_path))
print('总类别数据：', training_set.num_classes)
model = Yolo_Person2(training_set.num_classes)
pretrained_dict = torch.load(opt.pre_trained_model_path)
model_dict = model.state_dict()
pretrained_dict = {
k: v
for k, v in pretrained_dict.items() if k in model_dict
}
model_dict.update(pretrained_dict)
model.load_state_dict(model_dict)
print('Success Init BOT single class model')
else:
if opt.pre_trained_model_type == "model":
model = torch.load(opt.pre_trained_model_path,
map_location=lambda storage, loc: storage)
else:
model = Yolo(training_set.num_classes)
model.load_state_dict(
torch.load(opt.pre_trained_model_path,
map_location=lambda storage, loc: storage))
# The following line will re-initialize weight for the last layer, which is useful
# when you want to retrain the model based on my trained weights. if you uncomment it,
# you will see the loss is already very small at the beginning.
nn.init.normal_(list(model.modules())[-1].weight, 0, 0.01)
log_path = os.path.join(opt.log_path, "{}".format(opt.year))
if os.path.isdir(log_path):
shutil.rmtree(log_path)
os.makedirs(log_path)
writer = SummaryWriter(log_path)
if torch.cuda.is_available():
writer.add_graph(
model.cpu(),
torch.rand(opt.batch_size, 3, opt.image_size, opt.image_size))
model.cuda()
else:
writer.add_graph(
model, torch.rand(opt.batch_size, 3, opt.image_size,
opt.image_size))
# print('training_set.num_classes',training_set.num_classes)
criterion = YoloLoss(training_set.num_classes, model.anchors,
opt.reduction)
# criterion = YoloLoss(training_set.num_classes, model.anchors, opt.reduction)
optimizer = torch.optim.SGD(model.parameters(),
lr=1e-5,
momentum=opt.momentum,
weight_decay=opt.decay)
best_loss = 1e10
best_epoch = 0
model.train()
num_iter_per_epoch = len(training_generator)
for epoch in range(opt.num_epoches):
if str(epoch) in learning_rate_schedule.keys():
for param_group in optimizer.param_groups:
param_group['lr'] = learning_rate_schedule[str(epoch)]
for iter, batch in enumerate(training_generator):
image, label = batch
if torch.cuda.is_available():
image = Variable(image.cuda(), requires_grad=True)
else:
image = Variable(image, requires_grad=True)
optimizer.zero_grad()
logits = model(image)
# print('logits.shape',logits.shape)
# print('label.shape',len(label))
# print('label',label)
loss, loss_coord, loss_conf, loss_cls = criterion(logits, label)
loss.backward()
optimizer.step()
print(
"Epoch: {}/{}, Iteration: {}/{}, Lr: {}, Loss:{:.2f} (Coord:{:.2f} Conf:{:.2f} Cls:{:.2f})"
.format(epoch + 1, opt.num_epoches, iter + 1,
num_iter_per_epoch, optimizer.param_groups[0]['lr'],
loss, loss_coord, loss_conf, loss_cls))
writer.add_scalar('Train/Total_loss', loss,
epoch * num_iter_per_epoch + iter)
writer.add_scalar('Train/Coordination_loss', loss_coord,
epoch * num_iter_per_epoch + iter)
writer.add_scalar('Train/Confidence_loss', loss_conf,
epoch * num_iter_per_epoch + iter)
writer.add_scalar('Train/Class_loss', loss_cls,
epoch * num_iter_per_epoch + iter)
if epoch % opt.test_interval == 0:
model.eval()
loss_ls = []
loss_coord_ls = []
loss_conf_ls = []
loss_cls_ls = []
for te_iter, te_batch in enumerate(test_generator):
te_image, te_label = te_batch
num_sample = len(te_label)
if torch.cuda.is_available():
te_image = te_image.cuda()
with torch.no_grad():
te_logits = model(te_image)
batch_loss, batch_loss_coord, batch_loss_conf, batch_loss_cls = criterion(
te_logits, te_label)
loss_ls.append(batch_loss * num_sample)
loss_coord_ls.append(batch_loss_coord * num_sample)
loss_conf_ls.append(batch_loss_conf * num_sample)
loss_cls_ls.append(batch_loss_cls * num_sample)
te_loss = sum(loss_ls) / test_set.__len__()
te_coord_loss = sum(loss_coord_ls) / test_set.__len__()
te_conf_loss = sum(loss_conf_ls) / test_set.__len__()
te_cls_loss = sum(loss_cls_ls) / test_set.__len__()
print(
"Epoch: {}/{}, Lr: {}, Loss:{:.2f} (Coord:{:.2f} Conf:{:.2f} Cls:{:.2f})"
.format(epoch + 1, opt.num_epoches,
optimizer.param_groups[0]['lr'], te_loss,
te_coord_loss, te_conf_loss, te_cls_loss))
writer.add_scalar('Test/Total_loss', te_loss, epoch)
writer.add_scalar('Test/Coordination_loss', te_coord_loss, epoch)
writer.add_scalar('Test/Confidence_loss', te_conf_loss, epoch)
writer.add_scalar('Test/Class_loss', te_cls_loss, epoch)
model.train()
if te_loss + opt.es_min_delta < best_loss:
best_loss = te_loss
best_epoch = epoch
# torch.save(model, opt.saved_path + os.sep + "trained_yolo_voc")
torch.save(
model.state_dict(),
opt.saved_path + os.sep + "only_params_trained_yolo_bot")
torch.save(
model,
opt.saved_path + os.sep + "whole_model_trained_yolo_bot")
# Early stopping
if epoch - best_epoch > opt.es_patience > 0:
print(
"Stop training at epoch {}. The lowest loss achieved is {}"
.format(epoch, te_loss))
break
writer.export_scalars_to_json(log_path + os.sep + "all_logs.json")
writer.close()
def train(opt):
if torch.cuda.is_available():
torch.cuda.manual_seed(123)
else:
torch.manual_seed(123)
learning_rate_schedule = {"0": 1e-5, "5": 1e-4, "80": 1e-5, "110": 1e-6}
training_params = {
"batch_size": opt.batch_size,
"shuffle": True,
"drop_last": True,
"collate_fn": custom_collate_fn
}
test_params = {
"batch_size": opt.batch_size,
"shuffle": False,
"drop_last": False,
"collate_fn": custom_collate_fn
}
training_set = VOCDataset(opt.data_path, opt.year, opt.train_set,
opt.image_size)
training_generator = DataLoader(training_set, **training_params)
test_set = VOCDataset(opt.data_path,
opt.year,
opt.test_set,
opt.image_size,
is_training=False)
test_generator = DataLoader(test_set, **test_params)
model = Yolo(training_set.num_classes)
log_path = os.path.join(opt.log_path, "{}".format(opt.year))
if os.path.isdir(log_path):
shutil.rmtree(log_path)
os.makedirs(log_path)
writer = SummaryWriter(log_path)
if torch.cuda.is_available():
writer.add_graph(
model.cpu(),
torch.rand(opt.batch_size, 3, opt.image_size, opt.image_size))
model.cuda()
else:
writer.add_graph(
model, torch.rand(opt.batch_size, 3, opt.image_size,
opt.image_size))
criterion = YoloLoss(training_set.num_classes, model.anchors,
opt.reduction)
optimizer = torch.optim.SGD(model.parameters(),
lr=1e-5,
momentum=opt.momentum,
weight_decay=opt.decay)
best_loss = 1e10
best_epoch = 0
model.train()
num_iter_per_epoch = len(training_generator)
for epoch in range(opt.num_epoches):
if str(epoch) in learning_rate_schedule.keys():
for param_group in optimizer.param_groups:
param_group['lr'] = learning_rate_schedule[str(epoch)]
for iter, batch in enumerate(training_generator):
image, label = batch
if torch.cuda.is_available():
image = Variable(image.cuda(), requires_grad=True)
else:
image = Variable(image, requires_grad=True)
optimizer.zero_grad()
logits = model(image)
loss, loss_coord, loss_conf, loss_cls = criterion(logits, label)
loss.backward()
optimizer.step()
print(
"Epoch: {}/{}, Iteration: {}/{}, Lr: {}, Loss:{:.2f} (Coord:{:.2f} Conf:{:.2f} Cls:{:.2f})"
.format(epoch + 1, opt.num_epoches, iter + 1,
num_iter_per_epoch, optimizer.param_groups[0]['lr'],
loss, loss_coord, loss_conf, loss_cls))
writer.add_scalar('Train/Total_loss', loss,
epoch * num_iter_per_epoch + iter)
writer.add_scalar('Train/Coordination_loss', loss_coord,
epoch * num_iter_per_epoch + iter)
writer.add_scalar('Train/Confidence_loss', loss_conf,
epoch * num_iter_per_epoch + iter)
writer.add_scalar('Train/Class_loss', loss_cls,
epoch * num_iter_per_epoch + iter)
if epoch % opt.test_interval == 0:
model.eval()
loss_ls = []
loss_coord_ls = []
loss_conf_ls = []
loss_cls_ls = []
for te_iter, te_batch in enumerate(test_generator):
te_image, te_label = te_batch
num_sample = len(te_label)
if torch.cuda.is_available():
te_image = te_image.cuda()
with torch.no_grad():
te_logits = model(te_image)
batch_loss, batch_loss_coord, batch_loss_conf, batch_loss_cls = criterion(
te_logits, te_label)
loss_ls.append(batch_loss * num_sample)
loss_coord_ls.append(batch_loss_coord * num_sample)
loss_conf_ls.append(batch_loss_conf * num_sample)
loss_cls_ls.append(batch_loss_cls * num_sample)
te_loss = sum(loss_ls) / test_set.__len__()
te_coord_loss = sum(loss_coord_ls) / test_set.__len__()
te_conf_loss = sum(loss_conf_ls) / test_set.__len__()
te_cls_loss = sum(loss_cls_ls) / test_set.__len__()
print(
"Epoch: {}/{}, Lr: {}, Loss:{:.2f} (Coord:{:.2f} Conf:{:.2f} Cls:{:.2f})"
.format(epoch + 1, opt.num_epoches,
optimizer.param_groups[0]['lr'], te_loss,
te_coord_loss, te_conf_loss, te_cls_loss))
writer.add_scalar('Test/Total_loss', te_loss, epoch)
writer.add_scalar('Test/Coordination_loss', te_coord_loss, epoch)
writer.add_scalar('Test/Confidence_loss', te_conf_loss, epoch)
writer.add_scalar('Test/Class_loss', te_cls_loss, epoch)
model.train()
if te_loss + opt.es_min_delta < best_loss:
best_loss = te_loss
best_epoch = epoch
# torch.save(model, opt.saved_path + os.sep + "trained_yolo_voc")
torch.save(
model.state_dict(),
opt.saved_path + os.sep + "only_params_trained_yolo_voc")
torch.save(
model,
opt.saved_path + os.sep + "whole_model_trained_yolo_voc")
# Early stopping
if epoch - best_epoch > opt.es_patience > 0:
print(
"Stop training at epoch {}. The lowest loss achieved is {}"
.format(epoch, te_loss))
break
writer.export_scalars_to_json(log_path + os.sep + "all_logs.json")
writer.close()
def train(opt):
if torch.cuda.is_available():
torch.cuda.manual_seed(123)
else:
torch.manual_seed(123)
learning_rate_schedule = {"0": 1e-5, "5": 1e-4, "80": 1e-5, "110": 1e-6}
training_params = {
"batch_size": opt.batch_size,
"shuffle": True,
"drop_last": True,
"collate_fn": custom_collate_fn,
"num_workers": 8
}
test_params = {
"batch_size": opt.batch_size,
"shuffle": False,
"drop_last": False,
"collate_fn": custom_collate_fn,
"num_workers": 8
}
training_set = VOCDataset('data/norm_train_image/',
'data/train_labels.csv',
data_augmentation=True)
training_generator = DataLoader(training_set, **training_params)
test_set = VOCDataset('data/norm_train_image/',
'data/val_labels.csv',
data_augmentation=False)
test_generator = DataLoader(test_set, **test_params)
model = Yolo(1) #num_classes = 1
# model.load_state_dict(torch.load(opt.pre_trained_model_path))
# The following line will re-initialize weight for the last layer, which is useful
# when you want to retrain the model based on my trained weights. if you uncomment it,
# you will see the loss is already very small at the beginning.
nn.init.normal_(list(model.modules())[-1].weight, 0, 0.01)
log_path = os.path.join(opt.log_path, "{}".format(opt.year))
if os.path.isdir(log_path):
shutil.rmtree(log_path)
os.makedirs(log_path)
writer = SummaryWriter(log_path)
if torch.cuda.is_available():
writer.add_graph(
model.cpu(),
torch.rand(opt.batch_size, 3, opt.image_size, opt.image_size))
model.cuda()
else:
writer.add_graph(
model, torch.rand(opt.batch_size, 3, opt.image_size,
opt.image_size))
criterion = YoloLoss(training_set.num_classes, model.anchors,
opt.reduction)
optimizer = torch.optim.SGD(model.parameters(),
lr=1e-5,
momentum=opt.momentum,
weight_decay=opt.decay)
best_loss = 1e10
best_epoch = 0
model.train()
num_iter_per_epoch = len(training_generator)
for epoch in range(opt.num_epoches):
if str(epoch) in learning_rate_schedule.keys():
for param_group in optimizer.param_groups:
param_group['lr'] = learning_rate_schedule[str(epoch)]
epoch_start_time = time.time()
train_loss = 0
for iter, batch in enumerate(training_generator):
image, label = batch
if torch.cuda.is_available():
image = Variable(image.cuda(), requires_grad=True)
else:
image = Variable(image, requires_grad=True)
optimizer.zero_grad()
logits = model(image)
loss, loss_coord, loss_conf, loss_cls = criterion(logits, label)
# print('***loss***', loss)
loss.backward()
optimizer.step()
# print("Epoch: {}/{}, Iteration: {}/{}, Lr: {}, Loss:{:.2f} (Coord:{:.2f} Conf:{:.2f} Cls:{:.2f})".format(
# epoch + 1,
# opt.num_epoches,
# iter + 1,
# num_iter_per_epoch,
# optimizer.param_groups[0]['lr'],
# loss,
# loss_coord,
# loss_conf,
# loss_cls))
train_loss += float(loss.detach().cpu().data.numpy())
progress = (u"\u2588" * (int(
float(iter + 1) / num_iter_per_epoch * 40))).ljust(40, '.')
msg = '[%03d/%03d] %2.2f sec(s) |%s| Current Loss:%.2f' % (
iter + 1, num_iter_per_epoch,
(time.time() - epoch_start_time), progress, loss)
print(msg, end='', flush=True)
back = '\b' * len(msg)
print(back, end='', flush=True)
writer.add_scalar('Train/Total_loss', loss,
epoch * num_iter_per_epoch + iter)
writer.add_scalar('Train/Coordination_loss', loss_coord,
epoch * num_iter_per_epoch + iter)
writer.add_scalar('Train/Confidence_loss', loss_conf,
epoch * num_iter_per_epoch + iter)
writer.add_scalar('Train/Class_loss', loss_cls,
epoch * num_iter_per_epoch + iter)
if epoch % opt.test_interval == 0:
model.eval()
loss_ls = []
loss_coord_ls = []
loss_conf_ls = []
loss_cls_ls = []
for te_iter, te_batch in enumerate(test_generator):
te_image, te_label = te_batch
num_sample = len(te_label)
if torch.cuda.is_available():
te_image = te_image.cuda()
with torch.no_grad():
te_logits = model(te_image)
batch_loss, batch_loss_coord, batch_loss_conf, batch_loss_cls = criterion(
te_logits, te_label)
loss_ls.append(batch_loss * num_sample)
loss_coord_ls.append(batch_loss_coord * num_sample)
loss_conf_ls.append(batch_loss_conf * num_sample)
loss_cls_ls.append(batch_loss_cls * num_sample)
te_loss = sum(loss_ls) / test_set.__len__()
te_coord_loss = sum(loss_coord_ls) / test_set.__len__()
te_conf_loss = sum(loss_conf_ls) / test_set.__len__()
te_cls_loss = sum(loss_cls_ls) / test_set.__len__()
print(
"\nEpoch: {}/{}, Lr: {}, Train Loss:{:.2f} Val Loss:{:.2f} (Coord:{:.2f} Conf:{:.2f} Cls:{:.2f})"
.format(epoch + 1, opt.num_epoches,
optimizer.param_groups[0]['lr'],
train_loss / opt.num_epoches, te_loss, te_coord_loss,
te_conf_loss, te_cls_loss))
writer.add_scalar('Test/Total_loss', te_loss, epoch)
writer.add_scalar('Test/Coordination_loss', te_coord_loss, epoch)
writer.add_scalar('Test/Confidence_loss', te_conf_loss, epoch)
writer.add_scalar('Test/Class_loss', te_cls_loss, epoch)
model.train()
if te_loss + opt.es_min_delta < best_loss:
best_loss = te_loss
best_epoch = epoch
# torch.save(model, opt.saved_path + os.sep + "trained_yolo_voc")
torch.save(
model.state_dict(),
opt.saved_path + os.sep + "only_params_trained_yolo_voc")
torch.save(
model,
opt.saved_path + os.sep + "whole_model_trained_yolo_voc")
print("***Best Model Updated!***")
# Early stopping
if epoch - best_epoch > opt.es_patience > 0:
print(
"Stop training at epoch {}. The lowest loss achieved is {}"
.format(epoch, te_loss))
break
writer.export_scalars_to_json(log_path + os.sep + "all_logs.json")
writer.close()
def train(opt):
torch.cuda.manual_seed(123)
learning_rate_schedule = {"0": 1e-5, "5": 1e-4, "80": 1e-5, "110": 1e-6}
training_params = {
"batch_size": opt.batch_size,
"shuffle": True,
"drop_last": True,
"collate_fn": custom_collate_fn
}
test_params = {
"batch_size": opt.batch_size,
"shuffle": False,
"drop_last": False,
"collate_fn": custom_collate_fn
}
training_set = VOCDataset(opt.data_path, opt.year, opt.train_set,
opt.image_size)
training_generator = DataLoader(training_set, **training_params)
test_set = VOCDataset(opt.data_path,
opt.year,
opt.test_set,
opt.image_size,
is_training=False)
test_generator = DataLoader(test_set, **test_params)
model = Yolo(training_set.num_classes)
#nn.init.normal_(list(model.modules())[-1].weight, 0, 0.01)
model.cuda()
criterion = YoloLoss(training_set.num_classes, model.anchors,
opt.reduction)
optimizer = torch.optim.SGD(model.parameters(),
lr=1e-5,
momentum=opt.momentum,
weight_decay=opt.decay)
best_loss = 1e10
best_epoch = 0
model.train()
num_iter_per_epoch = len(training_generator)
for epoch in range(opt.num_epoches):
if str(epoch) in learning_rate_schedule.keys():
for param_group in optimizer.param_groups:
param_group['lr'] = learning_rate_schedule[str(epoch)]
for iter, batch in enumerate(training_generator):
image, label = batch
image = Variable(image.cuda(), requires_grad=True)
optimizer.zero_grad()
logits = model(image)
loss, loss_coord, loss_conf, loss_cls = criterion(logits, label)
loss.backward()
optimizer.step()
print(
"Epoch: {}/{}, Iteration: {}/{}, Lr: {}, Loss:{:.2f} (Coord:{:.2f} Conf:{:.2f} Cls:{:.2f})"
.format(epoch + 1, opt.num_epoches, iter + 1,
num_iter_per_epoch, optimizer.param_groups[0]['lr'],
loss, loss_coord, loss_conf, loss_cls))
if epoch % opt.test_interval == 0:
model.eval()
loss_ls = []
loss_coord_ls = []
loss_conf_ls = []
loss_cls_ls = []
for te_iter, te_batch in enumerate(test_generator):
te_image, te_label = te_batch
num_sample = len(te_label)
te_image = te_image.cuda()
with torch.no_grad():
te_logits = model(te_image)
batch_loss, batch_loss_coord, batch_loss_conf, batch_loss_cls = criterion(
te_logits, te_label)
loss_ls.append(batch_loss * num_sample)
loss_coord_ls.append(batch_loss_coord * num_sample)
loss_conf_ls.append(batch_loss_conf * num_sample)
loss_cls_ls.append(batch_loss_cls * num_sample)
te_loss = sum(loss_ls) / test_set.__len__()
te_coord_loss = sum(loss_coord_ls) / test_set.__len__()
te_conf_loss = sum(loss_conf_ls) / test_set.__len__()
te_cls_loss = sum(loss_cls_ls) / test_set.__len__()
print(
"Epoch: {}/{}, Lr: {}, Loss:{:.2f} (Coord:{:.2f} Conf:{:.2f} Cls:{:.2f})"
.format(epoch + 1, opt.num_epoches,
optimizer.param_groups[0]['lr'], te_loss,
te_coord_loss, te_conf_loss, te_cls_loss))
if te_loss + opt.es_min_delta < best_loss:
best_loss = te_loss
best_epoch = epoch
# torch.save(model, opt.saved_path + os.sep + "trained_yolo_voc")
torch.save(
model.state_dict(),
opt.saved_path + os.sep + "only_params_trained_yolo_voc")
torch.save(
model,
opt.saved_path + os.sep + "whole_model_trained_yolo_voc")
# Early stopping
if epoch - best_epoch > opt.es_patience > 0:
print(
"Stop training at epoch {}. The lowest loss achieved is {}"
.format(epoch, te_loss))
break